Table 1.

Model performance metrics on blind holdout test sets.

Dataset	Variant category	Method	AUC	MCC	TPR	TNR	PPV	NPV	Balanced accuracy	FP	TP	FN	TN	Call rate
Blind test set— BRCA SARC UCEC	Overall: 30,270 somatic, 31,359 germline	TabNet	0.942	0.762	0.931	0.828	0.839	0.926	0.879	5397	28,184	2086	25,962	100
		XGBoost	0.946	0.757	0.865	0.892	0.885	0.873	0.878	3392	26,186	4084	27,967	100
		LightGBM	0.949	0.766	0.874	0.892	0.886	0.88	0.883	3399	26,451	3819	27,960	100
		PureCN	0.85	0.592	0.662	0.912	0.882	0.729	0.787	2237	16,792	8588	23,065	82.2
	SNVs: 29,593 somatic, 30,122 germline	TabNet	0.945	0.775	0.943	0.827	0.843	0.937	0.885	5199	27,918	1675	24,923	100
		XGBoost	0.949	0.77	0.879	0.891	0.888	0.882	0.885	3286	26,006	3587	26,836	100
		LightGBM	0.951	0.776	0.884	0.892	0.889	0.887	0.888	3268	26,159	3434	26,854	100
		PureCN	0.851	0.593	0.662	0.913	0.887	0.724	0.788	2082	16,343	8360	21,983	81.7
	Indels: 839 somatic, 1762 germline	TabNet	0.784	0.433	0.838	0.624	0.515	0.89	0.731	662	703	136	1100	100
		XGBoost	0.799	0.408	0.666	0.757	0.566	0.827	0.712	428	559	280	1334	100
		LightGBM	0.815	0.455	0.765	0.718	0.564	0.865	0.742	497	642	197	1265	100
		PureCN	0.827	0.529	0.606	0.893	0.756	0.805	0.749	132	410	267	1105	73.6
Blind test set —metastatic melanoma	Overall: 15,813 somatic, 12,871 germline	TabNet	0.852	0.55	0.797	0.753	0.799	0.751	0.775	3176	12,598	3215	9695	100
		XGBoost	0.861	0.558	0.801	0.756	0.802	0.756	0.779	3135	12,670	3143	9736	100
		LightGBM	0.867	0.57	0.823	0.744	0.798	0.774	0.784	3289	13,020	2793	9582	100
		PureCN	0.824	0.52	0.789	0.732	0.794	0.726	0.76	2951	11,340	3041	8050	88.5
	SNVs: 15,688 somatic, 12,412 germline	TabNet	0.85	0.545	0.796	0.749	0.8	0.744	0.773	3114	12,490	3198	9298	100
		XGBoost	0.862	0.565	0.815	0.748	0.804	0.762	0.782	3125	12,788	2900	9287	100
		LightGBM	0.868	0.573	0.833	0.737	0.8	0.777	0.785	3266	13,061	2627	9146	100
		PureCN	0.823	0.516	0.788	0.729	0.797	0.717	0.758	2856	11,228	3028	7686	88.2
	Indels: 147 somatic, 611 germline	TabNet	0.82	0.361	0.837	0.619	0.346	0.94	0.728	233	123	24	378	100
		XGBoost	0.833	0.384	0.837	0.646	0.363	0.943	0.742	216	123	24	395	100
		LightGBM	0.84	0.343	0.898	0.534	0.317	0.956	0.716	285	132	15	326	100
		PureCN	0.887	0.624	0.888	0.826	0.581	0.964	0.857	80	111	14	379	77

AUC area under the receiver operating characteristic curve, MCC Matthews Correlation Coefficient, TP true positives— somatic mutations correctly classified as somatic, FP false positives—rare germline variants misclassified as somatic mutations, FN false negatives—somatic mutations misclassified as germline variants, TN true negatives—rare germline mutations correctly classified as germline. Bold values indicate the best performance for a given metric, variant category, and test set.

Benchmark accuracy metrics for tumor-only somatic vs germline classification by TabNet, XGBoost, LightGBM, and PureCN on blind test datasets. Overall performance considers all single-nucleotide variants (SNVs) and indels.

Call rate—percentage of total coding variants classified.